Open World Object Detection (OWOD) faces a fundamental dilemma: maintaining a stable representation for known classes while reserving flexible space for discovering unknown objects. Existing methods, while improving recall, often fail to assign discriminative confidence scores to unknown instances, resulting in critically low Average Precision (U-AP) and representation degradation during incremental learning. To remedy this, we propose the Evidential Prior Guided Neural Collapse (ENC) framework. ENC unifies representation learning and uncertainty quantification via a Geometric-Evidence Coupling mechanism. Unlike previous approaches, we map evidential support directly to the angular alignment with Simplex Equiangular Tight Frame (ETF) prototypes. Theoretically, the evidential prior functions as a geometric regularizer: it maximizes equiangular separation for confident known samples, while constraining ambiguous queries to approximate an isotropic uniform distribution via distributional regularization. Furthermore, to mitigate decision conflicts in self-supervised learning, we propose a dissonance-aware objectness optimization strategy that mines informative samples near the decision boundary. Extensive experiments on M-OWODB and S-OWODB benchmarks demonstrate that ENC sets a new state-of-the-art. Notably, it achieves a significant improvement in unknown class discovery, boosting U-AP from ≈ 1% to 9.2%, while exhibiting superior robustness against catastrophic forgetting in challenging incremental scenarios.

Brain tumors are highly lethal and debilitating pathological changes that require timely diagnosis and treatment. Magnetic resonance imaging (MRI), a non-invasive diagnostic tool, provides complementary multi-modal information crucial for accurate tumor detection and delineation. However, existing methods struggle to effectively fuse multi-modal information from MRI sequences and often fail to perform modality-specific feature extraction, which hinders accurate tumor segmentation. Furthermore, the inherent challenges posed by the blurred boundaries and complex morphological characteristics of tumor structures present additional substantial obstacles to achieving precise segmentation. To address these issues, we propose FiHam, a fine-grained hierarchical progressive modal-aware network that introduces a novel multi-modal fusion strategy and an advanced feature extraction mechanism. Specifically, FiHam employs a progressive fusion strategy that extracts modality-specific features at lower levels and integrates multi-modal features at higher levels to effectively leverage complementary information from tumor images. Additionally, we design a gated cross-attention modal-fusion module that adaptively selects and integrates dual-modal features using cross-attention mechanisms to enhance modality fusion. To further refine segmentation accuracy, we incorporate a tiny U-Net into the encoder to capture boundary features and complex tumor morphology. Extensive experiments on three large-scale, multi-modal brain tumor datasets demonstrate that FiHam achieves state-of-the-art performance, delivering significant improvements in segmentation accuracy and generalizability across diverse MRI modalities.

Optical Coherence Tomography (OCT) imaging is extensively utilized for non-invasive observation of pathological conditions, such as retinal fluid-associated diseases. Accurate fluid segmentation in OCT images is therefore critical for quantifying disease severity and aiding clinical decision-making. However, achieving precise segmentation remains challenging due to pathological variations in shape and size, uncertain boundaries, and low contrast of fluid. Most importantly, variability in OCT image styles across different vendors and centers significantly affects fluid segmentation, leading to poor generalization to unseen domains. To address this, we propose a novel method, RSAPower, to enhance the generalization ability of fluid perception networks via style augmentation for retinal fluid segmentation. Specifically, RSAPower comprises a plug-and-play random style transform augmentation (RSTAug) module and a novel fluid perception network (FLPNet) for end-to-end training. The RSTAug module generates new random-style data from the source domain, preserving realistic pathological and structural features. The FLPNet benefits from a novel hybrid structure attention (HSA) module to perceive fluid's spatial features and long-range dependence. Furthermore, FLPNet adapts to the diverse augmented data through a saliency-guided multi-scale attention (SGMA) block, boosting its segmentation performance. We validate RSAPower against various state-of-the-art methods using two publicly available datasets, Retouch and Kermany. Experimental results demonstrate the proposed method's superior generalization ability and effectiveness in fluid segmentation.

Synthetic aperture radar (SAR) images have limited labeled samples, and thus, it is difficult to learn a perfect convolutional neural network (CNN) model for target classification. The commonly used single-channel SAR images have much less information than those of the three-channel natural images. Transfer learning (TL) is an effective way to improve the generalization ability of the CNN model. The existing TL methods for SAR images usually transfer the knowledge from the three-channel natural images to the single-channel SAR images, where the SAR images are simply duplicated from one channel to three channels. This is obviously not reasonable. Indeed, the single-channel SAR image is complex valued, which can be divided into multiple channels (e.g., three channels) via the subaperture decomposition (SD) algorithm. In order to fully utilize the complex-valued data of the single-channel SAR images, in this letter, we propose a novel TL method with SD (TL-SD), where the SD can generate pseudocolor SAR images to realize TL with the large-scale natural image data sets. The experimental results based on the MSTAR real data set show that the proposed TL-SD method achieves an average accuracy of 99.88% on classification of ten-class targets and is superior to the other compared target classification methods, which verify the effectiveness of the proposed method.

Traditional ship detection methods for synthetic aperture radar (SAR) mainly utilize the amplitude information to distinguish ship targets from sea clutter, including constant false alarm rate (CFAR), visual attention model, and deep learning methods. The CFAR algorithms adopt the dense sliding window strategy, which is very time-consuming and may generate numerous false alarms. The deep learning methods are supervised and difficult to obtain satisfactory performance when the number of labeled samples is insufficient. The visual attention models can quickly focus on the potential target area, and however, it is still difficult to eliminate the strong clutter, such as radio frequency interference and azimuth ambiguity. In fact, as a coherent imaging system, SAR data itself are complex-valued. Compared with the amplitude information, complex information can essentially reflect the difference between ship target and sea clutter. To improve the accuracy and efficiency of ship detection, in this letter, a novel unsupervised ship detection method based on multiscale saliency and complex signal kurtosis (MSS-CSK) for single-channel SAR images is proposed, which contains the proposal extraction stage and the target discrimination stage. The experimental results based on the Radarsat-2 real SAR data show that the proposed method has high detection accuracy and efficiency.

To solve the imperfect springs structure parameters in the design of energy storage springs of the universal circuit breakers, and problems such as large volume of circuit breakers and low design efficiency, an approach to optimize the parameters of the energy storage springs of the circuit breakers is proposed based on the Artificial Bee Colony (ABC) algorithm. First, the mathematical optimization model of energy storage springs and the constraints of the spring parameters are derived in accordance with the working principle of energy storage springs. Then combined with cloud model and cross operation, the ABC algorithm is improved, which can adjust the cross factor, accelerate the convergence speed of ABC algorithm and improve the global search ability. And the classical test-function simulations verify that the improved ABC algorithm is superior to other evolutionary algorithms. Finally, the two different types of energy storage springs optimization models of universal circuit breakers are experimentally analyzed by use of the improved ABC algorithm, and the corresponding springs’ parameters are calculated. The experiment results show that the proposed approach is effective, which can reasonably design the parameters of energy storage springs of circuit breakers and improve the design efficiency.

The traditional robust principal component analysis (RPCA) model is based on the nuclear norm, which usually underestimates the singular values of the low‐rank matrix. As a consequence, the restoration image experiences serious interference by Gaussian noise, and the image quality degenerates during the denoising process. Therefore, an improved RPCA method via non‐convex regularisation terms is proposed to remedy the above shortcomings. First, in order to estimate the singular value of the low‐rank matrix more accurately, the authors employ the non‐convex penalty function and add a weight vector to it. Then, the regularisation with non‐convex penalty function and its weighted version are used to replace the nuclear norm and entry‐wise norm in original RPCA, respectively, to establish an improved model. Finally, an optimal solution algorithm is derived by developing the alternating direction multiplier method. Experimental results show that the proposed method has better performance in terms of both quantitative measurement and visual perception quality than other several state‐of‐the‐art image denoising methods.

A novel slow-wave structure optimization method on Chaos-improved Normal mutation cat swarm optimization (CI-NMCSO) algorithm is proposed. Under the variable helix section length in STWT, the CI-NMCSO combined with 1D CHRISTINE code is used to calculate the best set of pitch distribution and section length with the objective function of electron beam efficiency improvement. Quantum particle swarm optimization (QPSO) and Cauchy mutated cat swarm optimization (CMCSO) algorithms are applied to make performance comparison. Experimental results show that the beam efficiency has been increased by CI-NMCSO from rated value 30% to 45.3%, and the values using CMCSO and QPSO are 41.8% and 36.5%, respectively, the convergence speed of CI-NMCSO is the fastest, only 16 iterations, while CMCSO and QPSO take 19 and 23 iterations, so the performance of CI-NMCSO is better than CMCSO and QPSO on both optimization precision and calculation speed in terms of slow-wave structure optimization, and is also superior to that with equal section length when the helix section length is variable.

Modelling users’ dynamic preference for personalized temporal recommendation has been a hot research topic. Traditional dynamic recommendation models divide a user’s interaction history into fixed-sized windows to learn the user’s evolving preference. This strategy however worsens the issue of data sparsity in that some sessions may have very little or even no interaction for preference inference. To alleviate the data sparsity issue and avoid errors due to data imputation that is commonly adopted by existing models, a novel session-based dynamic recommendation model that divides a user’s interaction history with dynamic window size is proposed. An empirical study on the users’ activity life cycles using real-world dataset is conducted to demonstrate the nonuniformness and aggregation of the users’ behavior patterns on the time dimension. Based on the study, a user’s interaction history is divided with dynamic temporal window size by using dynamic clustering. The user’s evolving profile is then constructed by modeling her preference in each session using Latent Dirichlet Allocation (LDA) and a time sensitive weighting scheme. Our dynamic model is designed for two major recommendation tasks: (1) top-N recommendation, which is provided by measuring the relevance of probabilistic topic distribution between the user’s profile in the next temporal domain and each candidate item, (2) rating prediction that is achieved by finding K nearest user/item neighbors based the similarity of probabilistic topic distribution between users/items. Extensive empirical experiments over two real datasets demonstrate the effectiveness and superiority of our method by comparing to representative temporal dynamic methods.